In the master's thesis we explore the problem of characterization and recognition of sounds on the valves of heating systems. We compare noise on the valves of various dimensions, functions and geometry. Based on the data base of soundtracks, suitable noise categories are formed. We extract statistical features in the time and frequency domains, with the help of statistical analysis of sound signals measured on valves. Based on the extracted features, we propose means to classify the sound signals into proper categories. With the help of decision thresholds we develop several classifying models formulated as decision trees. The sequence of nodes and value of decision thresholds are generated manually and numerically. Our requirements are a simple decision tree, low classification loss and a good separation of the phenomena of squealing and cavitation. According to our requirements we choose two models that are most suitable for practical use. First is decision tree with branching level 11, which has the lowest classification loss. Second is decision tree with manually defined structure and numerically defined decision thresholds. The latter best distinguishes phenomena of squealing and cavitating.
|
